S3R: Modeling spatially varying associations with Spatially Smooth Sparse Regression

Area of Science:

• Genomics
• Computational Biology
• Statistical Modeling

Background:

• Spatial transcriptomics (ST) data presents challenges like noise, cell mixing, and high dimensionality.
• Existing models struggle to capture dynamic molecular associations across tissue locations.

Purpose of the Study:

• To introduce Spatially Smooth Sparse Regression (S3R), a novel statistical framework for analyzing ST data.
• To develop a method that estimates location-specific coefficients for linking molecular features across tissue space.
• To provide a scalable and interpretable regression framework for diverse biological questions in ST.

Main Methods:

• S3R integrates structured sparsity with a minimum-spanning-tree-guided smoothness penalty.
• The framework estimates location-specific coefficients for high-dimensional spatial predictors.
• An efficient implementation is provided to handle large ST datasets.

Main Results:

• S3R accurately recovers spatially varying effects and selects relevant predictors in synthetic data.
• It recapitulates layer-specific gene associations in human brain ST data.
• S3R demixes cell type-attributed expression fields in infection and cancer data, revealing spatial gradients and cell-cell interactions.
• Analysis of breast cancer data delineates gene expression contributions at multiple contextual levels.

Conclusions:

• S3R offers a robust and flexible approach to dissecting complex spatial relationships in ST data.
• The method enhances the biological interpretability of gene expression patterns and cell-cell crosstalk.
• S3R provides a scalable and unified framework for addressing various ST analysis challenges.